tqchen opened a new pull request, #619:
URL: https://github.com/apache/tvm-ffi/pull/619
The object-graph deserializer (`FromJSONGraph`) trusted several values
taken directly from the input JSON, which let a malformed or adversarial
graph trigger out-of-bounds reads, a stack overflow, or a segfault
instead of a clean error.
## Changes (`src/ffi/extra/serialization.cc`)
- Validate every node index (`root_index` and all child/key/value/field
references) against the nodes array before indexing into the
`decoded_nodes_` `std::vector`. Previously a negative index segfaulted
in `GetOrDecodeNode` and an out-of-range index read past the vector end.
- Add a recursion-depth guard so a deeply nested graph or a long chain of
node references throws rather than overflowing the C++ stack; this also
makes a self-referential node terminate by throwing.
- Require `Device` data to be exactly `[device_type, device_id]`.
- Require `Map`/`Dict` data to have an even number of `[key, value]`
entries so the inner loop cannot read one element past the end.
All previously-trusted indices, sizes, kinds and keys now throw an
`ffi::Error` (`ValueError` / `TypeError` / `KeyError` / `IndexError`)
before any unsafe access. Type-key, missing-key, and wrong-value-type
cases were already routed through checked accessors and continue to throw.
## Tests
- `tests/cpp/extra/test_serialization.cc`: a battery of crafted malformed
graphs (out-of-range / negative indices, odd-length map data, wrong
device arity, unknown type keys, missing keys, wrong value types,
non-object nodes, truncated JSON, deeply nested input, self-cycle), each
asserted to throw. The C++ test target builds with AddressSanitizer and
the full suite passes clean; the prior unhardened code segfaults on the
negative-index case under the same build.
- `tests/python/test_serialization.py`: the same cases at the Python
boundary, each asserted to raise.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
